Avoid creating illegal byref pointers (#17524)
Byref pointers need to point within their "host" object -- thus
the alternate name "interior pointers". If the JIT creates and
reports a pointer as a "byref", but it points outside the host
object, and a GC occurs that moves the host object, the byref
pointer will not be updated. If a subsequent calculation puts
the byref "back" into the host object, it will actually be pointing
to garbage, since the host object has moved.
This occurred on ARM with array index calculations, in particular
because ARM doesn't have a single-instruction "base + scale*index + offset"
addressing mode. Thus, we were generating, for the jaggedarr_cs_do
test case, `ProcessJagged3DArray()` function:
```
// r0 = array object, r6 = computed index offset. We mark r4 as a byref.
add r4, r0, r6
// r4 - 32 is the offset of the object we care about. Then we load the array element.
// In this case, the loaded element is a gcref, so r4 becomes a gcref.
ldr r4, [r4-32]
```
We get this math because the user code uses `a[i - 10]`, which is
essentially `a + (i - 10) * 4 + 8` for element size 4. This is optimized
to `a + i * 4 - 32`. In the above code, `r6` is `i * 4`. In this case,
after the first instruction, `r4` can point beyond the array.
If a GC happens, `r4` isn't updated, and the second instruction loads garbage.
There are several fixes:
1. Change array morphing in `fgMorphArrayIndex()` to rearrange the array index
IR node creation to only create a byref pointer that is precise; don't create
"intermediate" byref pointers that don't represent the actual array element
address being computed. The tree matching code that annotates the generated tree
with field sequences needs to be updated to match the new form.
2. Change `fgMoveOpsLeft()` to prevent the left-weighted reassociation optimization
`[byref]+ (ref, [int]+ (int, int)) => [byref]+ ([byref]+ (ref, int), int)`. This
optimization creates "incorrect" byrefs that don't necessarily point within
the host object.
3. Add an additional condition to the `Fold "((x+icon1)+icon2) to (x+(icon1+icon2))"`
morph optimization to prevent merging of constant TYP_REF nodes, which now were
being recognized due to different tree shapes. This was probably always a problem,
but the particular tree shape wasn't seen before.
These fixes are all-platform. However, to reduce risk at this point, the are
enabled for ARM only, under the `FEATURE_PREVENT_BAD_BYREFS` `#ifdef`.
Fixes #17517.
There are many, many diffs.
For ARM32 ngen-based desktop asm diffs, it is a 0.30% improvement across all
framework assemblies. A lot of the diffs seem to be because we CSE the entire
array address offset expression, not just the index expression.